LTX Meets NVLink: Why You Can't Split One Render Across Macs

Everyone asks the same question about a Mac render farm: can I gang four Macs together so one clip renders four times faster?

No. And the reason is a single number, so let's do it properly instead of hand-waving.

The number

Diffusion is iterative. A 97-frame clip is ~40 denoising steps, and a tensor-parallel split means every participating machine exchanges activation tensors at every step. Inside a DGX that's NVLink at ~900 GB/s. Between your Macs it's this:

NVLink            ~900   GB/s
Thunderbolt 4      ~5    GB/s   (and not a fabric)
10GbE              ~1.25 GB/s
Gigabit ethernet   ~0.125 GB/s   <-- your switch

900 / 0.125 = 7,200× less bandwidth

Even generously, gigabit is three to four orders of magnitude short. Add per-step latency and the “cluster” finishes after a single Mac would have. You built a distributed system to go slower. Do not @ me — measure it.

The parallelism that does work

Stop splitting the render. Split the work. One clip per Mac, no coordination during the render at all:

queue/  a.job  b.job  c.job  d.job      # 1KB text files
           |      |      |      |
         Mac1   Mac2   Mac3   Mac4      # each claims one with mv
           |      |      |      |
done/    a.mp4  b.mp4  c.mp4  d.mp4      # only these cross the wire

Nothing touches the network between claim and finish. Four Macs = ~4× clips per hour, near-linear, and the interconnect stops mattering — which is exactly why a $20 switch is enough. That's the model the whole farm is built on.

“But my batch is one clip”

Then the farm can't help you, and neither can honesty-free marketing. What the farm can do is make one clip cheaper to get right: render a batch of cheap proof stills across all four Macs in seconds, cherry-pick the framing, and spend one Mac's hour on the winner instead of guessing.

# 4 seeds of a proof still, split across the farm
./enqueue.sh --id roofline --sweep 4 --test \
  --prompt "storm clouds over a QLD tile roof at dusk"

That's the same trick a colourist uses with contact sheets, and it's worth more than a fake 4× on a single render.

The catch

Job-level parallelism has a floor: your slowest single render. If one hero clip takes 30 minutes, no amount of Macs makes that clip arrive sooner. Plan around throughput, not latency — queue the night, not the next five minutes.

The verdict

NVLink is a fabric; ethernet is a mailbox. Design for the mailbox and Apple Silicon is genuinely useful: unified memory means a 64GB Mac holds models a 24GB GPU can't, and you own the hardware outright. Design for the fabric you don't have and you'll ship a slower renderer with more moving parts.